Detect. Match. Score. Reroute.
The risk engine sits beside your load balancer, never in front of it. It reads public vulnerability data and your signed inventories, decides, and writes weights through your load balancer's own config API.
Detect
Pull OSV.dev, CISA KEV, FIRST EPSS, distro and FreeBSD advisories, and signed VEX statements. Feeds are verified and time-stamped; stale or unsigned data freezes changes.
Match
Compare each advisory with a signed SBOM tied to the image that is actually deployed, cross-checked by an independent scan. Never the Server header, never just what's in Git.
Score
Known exploited? Likely to be exploited? Reachable in your build? Severity alone doesn't move traffic.
Check
Capacity guard (never drain every stack), hysteresis and dwell timers (no flapping), and your approval policy for hard cuts.
Reroute
Write new backend weights through the load balancer's config API. GCP Global Load Balancer first; AWS, Azure and Cloudflare are planned.
Verify
Confirm the change landed and the healthy stack is carrying load. If not, roll back to the last known-good weights.
Report
Every decision is logged: feed entry, inventory match, score, guard results, approver and outcome.
Customer traffic in blue. Our control plane in orange.
The only thing the engine touches is load balancer configuration. If Doppelstack goes dark, your load balancer keeps serving with its last weights.
Optional stay-up mode: if every stack is hit at once, the load balancer can fail over to a read-only static copy of your site, so customers see something useful instead of an error.

The second stack is the other half
Least-vulnerable routing needs somewhere safe to send traffic. If you only run one implementation today, Stack Duplication builds its twin.